The web consoles
Spancache has two web surfaces behind one sign-in. The workspace (api.spancache.ai) is
where you set up your account — ingest token, usage, plan, team. The trace console
(app.spancache.ai) is where you explore the data your agents send. A single domain-wide session
covers both, so you move between them without signing in twice.
A new user starts in the workspace to grab a token and point an exporter at the endpoint (see getting-started), then crosses into the console the moment traces land.
The workspace — api.spancache.ai
Account administration. Nothing here searches your data; it’s the control surface around it.
| Area | What it’s for | What you do |
|---|---|---|
| Ingest token | The per-tenant bearer token that authorizes your traces. | Copy it into your exporter’s Authorization: Bearer … header. Reveal, rotate, or revoke it here — rotating invalidates the old one. |
| Usage | What you’ve sent and how it’s stored. | Watch spans ingested, stored bytes, and the compression ratio (raw ÷ stored) over the current period, with a daily trend. This is also the metered basis for your bill. |
| Plan & billing | Your capacity tier and payment. | See the tier your measured ingest and retention land in, compare tiers, and manage the payment method. Cost detail lives in cost. |
| Team | Who can see the workspace. | Invite members, assign roles, and remove access. Invited members activate from an emailed link. |
| Capture level | Metadata-first vs. content capture. | Spancache records timing, model, tokens, and cost by default and no prompt/response text. Turn on content capture here; it’s scrubbed for secrets at ingest before anything is stored. |
A prominent Open Search → button jumps straight to the trace console, carrying your session with it.
The trace console — app.spancache.ai
Where you read the data. Five tabs, all built on the same query engine.
Traces
Recent agent runs, one row per trace: service, span count, duration, total cost, error count, and start time. Click a row to open the span waterfall — the tree of every step in that run, LLM calls and tool calls nested by parent, each with its own cost, latency, and status. If content capture is on, a span expands to show the prompt and response for that call (secrets already scrubbed). This is the view for answering “what did this run actually do, and where did the time and money go?”
Cost
Spend rolled up two ways: per model and per release version. Each row shows spend, call count, error rate, p95 latency, and tokens, so a cost or latency regression between versions — a new prompt that doubled tokens, a model swap that raised the error rate — shows up as a moved number rather than a surprise on the invoice. Cost is computed from tokens × list price, cache-aware, on every span; the method is in cost.
Detections
One-click guardrail templates for the common agent failure modes, each a saved rule you enable without writing a query:
- Runaway cost — a run’s spend crosses a threshold.
- Tool-error loop — the same tool keeps failing within a run.
- Error-rate spike — failures climb across runs.
- Output truncation — responses cut off at the token limit.
- Retry storm — a step retries far past normal.
- Slow first token — time-to-first-token regresses.
- Context burn — prompt tokens balloon toward the window.
- Slow agent step — a single span’s latency crosses a threshold.
Enabled rules are evaluated exactly across the whole fleet, and what they fire lands in an alert triage inbox on the same tab, where you acknowledge and work through them.
Anomalies
What the templates don’t know to ask for. Spancache surfaces cost-outlier traces (runaway loops, cache misses that should have hit), novel agent behavior (a run shaped unlike anything seen before), and incompressible output — text that resists compression, a signal of possible prompt-injection or data exfil. These are found automatically, no rule to configure.
Search
The raw query surface everything else is built on: a query bar, a results table, and a field
explorer that shows the top fields and values so you can pivot with a click. Write
<search> | stats … by … to slice traces and spans any way the pre-built views don’t. See
query.
Moving between the two
The surfaces cross-link both directions: Open Search → takes you from the workspace into the trace console, and Workspace ↗ takes you back from the console to account and billing. One login via a domain-wide session means neither hop asks you to sign in again.